Targeting Alignment: Extracting Safety Classifiers of Aligned LLMs